Neural net controller for noise and vibration reduction

• US 6,493,689 B2
• Filed: 12/29/2000
• Issued: 12/10/2002
• Est. Priority Date: 12/29/2000
• Status: Expired due to Term

• Current Assignee: General Dynamics Mission Systems Inc. (General Dynamics Corporation)
• Original Assignee: General Dynamics Advanced Technology Systems, Inc. (General Dynamics Corporation)
• Inventors: Torok, Michael S., Berezin, Charles, Lorber, Peter F., Kotoulas, Antonios N.
• Primary Examiner(s): Black, Thomas
• Assistant Examiner(s): Holmes, Michael B.
• Application Number:
  US09/752,180
• Publication Number:
  US 20020117579A1
• Time in Patent Office: 711 Days
• Field of Search: 706/23, 608/03, 700/28, 700/280
• US Class Current: 706/23
• CPC Class Codes: B64C 2027/7266 actuated by actuators
• CPC Class Codes: F16F 15/02 Suppression of vibrations of non-rotating, e.g. reciprocating systems; Suppression of vibrations of rotating systems by use of members not moving with the rotating systems F16F15/005 takes precedence ; layered products B32B; suppression of vibration in ships B63; relieving load on bearings, using magnetic means F16C39/06
• CPC Class Codes: F16F 2230/18 Control arrangements
• CPC Class Codes: G10K 11/178 by electro-acoustically regenerating the original acoustic waves in anti-phase
• CPC Class Codes: Y02T 50/30 Wing lift efficiency

Abstract

Two neural networks are used to control adaptively a vibration and noise-producing plant. The first neural network, the emulator, models the complex, nonlinear output of the plant with respect to certain controls and stimuli applied to the plant. The second neural network, the controller, calculates a control signal which affects the vibration and noise producing characteristics of the plant. By using the emulator model to calculate the nonlinear plant gradient, the controller matrix coefficients can be adapted by backpropagation of the plant gradient to produce a control signal which results in the minimum vibration and noise possible, given the current operating characteristics of the plant.

Claims

1. A system for actively controlling noise and vibration produced by the flow of a fluid over a surface as measured by a plurality of noise and vibration sensors by altering the flow of the fluid, said system comprising:

a plant comprising at least a portion of a plurality of mechanical equipment involved in producing and measuring the noise and vibration;

a preprocessing filtering module which isolates and quantifies the signals from the sensors at least one of which signals represent the measured noise and vibration desired to be controlled;

means for measuring at least one stimulus related to a targeted noise and vibration;

a reference signal;

an adaptive controller neural network which receives an input from said reference signal and produces an adjustment signal;

a means for altering the flow of the fluid by the state of said altering means and which receives as input said adjustment signal;

an emulator neural network trained to model the dynamics of the fluid flow, said model receiving as input at least one of said stimulus measurements and the state of said altering means, calculating estimated noise and vibration signals, comparing the estimates with the results from the sensor preprocessing filtering module, producing a plant gradient signal representing an error in the model and feeding back said plant gradient signal into said controller neural network to adapt aid controller neural network; and

a plurality of time-based filters which produces time-based representations of the inputs to said emulator neural network and said controller neural network.

2. The system according to claim 1 wherein said emulator neural network is a feed-forward multi-layer neural network further comprising:

an input vector with dimension J comprising at least one of the values stored in said time-based filters at the input to said emulator;

a first layer having a number of neurons equal to (Q+1);

an output layer having a number of neurons L+1 wherein (L+1) is equal to number said noise sensors plus the number said vibration sensors;

at least one intermediate layer having a number of neurons M+1;

a plurality of connections w_j,q⁽¹⁾ between the input vector and the neurons in the first layer, w_q,m⁽²⁾, between the neurons in any successive intermediate layers, and w_m,l⁽³⁾ between the neurons in the last intermediate layer and said output layer;

a modeling section for calculating an estimated result from each of the input, output and intermediate neurons;

a memory section for storing a plurality of prior estimate results from the neurons in said input layer, said output layer and each of said intermediate layers;

a performance section for calculating the error gradient of the plant during a backward pass of said measured noise and vibration; and

an update section for updating the plant gradients to said controller network.

3. The system according to claim 1 wherein said controller neural network is a feed-forward multi-layer neural network comprising:

an input vector with dimension J comprising at least one of the values stored in said time-based filters at the input to said controller;

a first layer having a number of neurons (K+1);

an output layer having a number of neurons L+1 equal to number of means for altering the fluid flow;

one or more intermediate layers having a number of neurons P+1;

a plurality of connections W_j,k^(1c) between the input vector and the neurons in the

first layer, w_k,p^(2c), between the neurons in any successive intermediate layers, and w_p,l^(3c) between the neurons in the last intermediate layer and said output layer;

an objective function section for calculating a desired state of the means for altering the flow of liquid;

a gradient descent performance section for adapting the connections {w_j,k^(1c), w_k,p^(2c),w_p,l^(3c)} according to the plant gradients received from the emulator network; and

an update section for updating the state of the controlled altering means to the altering means and to the input of the emulator network.

4. The system according to claim 2 wherein said controller neural network is a feed forward multi-layer neural network comprising:

an input vector with dimension J comprising at least one of the values stored in said time-based filters at the input to said controller;

a first layer having a number of neurons (K+1);

an output layer having a number of neurons L+1 equal to number of means for altering the fluid flow;

one or more intermediate layers having a number of neurons P+1;

a plurality of connections w_j,k^(1c) between the input vector and the neurons in the first layer, w_k,p^(2c), between the neurons in any successive intermediate layers, and w_p,l^(3c) between the neurons in the last intermediate layer and said output layer;

an objective function section for calculating a desired state of the means for altering the flow of liquid;

a gradient descent performance section for adapting the connections {w_j,k^(1c), w_k,p^(2c),w_p,l^(3c)} according to the plant gradients received from the emulator network; and

an update section for updating the state of the controlled altering means to the altering means and to the input of the emulator network.

5. The system according to claim 4 wherein said time-based filter comprises memorydelay lines which store the input values for a given time period and permit access to each of the stored values.

6. The system according to claim 5 wherein said state of altering means input signal to said emulator networks comprises said adjustment signal from said controller network.

7. The system according to claim 6 wherein said surface is being driven through said fluid.

8. The system according to claim 7 wherein said driven surface is attached to a shaft and driven rotationally through said fluid.

9. The system according to claim 8 further comprising means for measuring the rotational rate at which said shaft is driven with said reference signal comprising harmonics of the rotational rate at which the surface is driven through the fluid.

10. The system according to claim 9 wherein said harmonics include the harmonic at integer multiples of the rotational rate.

11. The system according to claim 10 wherein said integer multiples include twice the rotational rate.

12. The system according to claim 8 further comprising means for measuring the axial position of said surface rotating around said shaft wherein at least one of said stimuli comprises said measured axial position of said surface rotating around said shaft.

13. The system according to claim 8 further comprising means for measuring the axial position of said surface rotating around said shaft wherein at least one of said stimuli comprises the sine of said measured axial position of said surface rotating around said shaft.

14. The system according to claim 13 further comprising means for measuring the axial position of said surface rotating around said shaft wherein at least one of said stimuli comprises said cosine of said measured axial position of said surface rotating around said shaft.

15. The system according to claim 14 further comprising means for measuring the forward velocity of a body attached to said surface through said fluid wherein at least one of said stimuli comprises said measured forward velocity.

16. The system according to claim 8 further comprising means for measuring the forward velocity of a body attached to said surface through said fluid wherein at least one of said stimuli comprises said measured forward velocity.

17. The system according to claim 16 further comprising means for measuring the axial position of said surface rotating around said shaft wherein at least one of said stimuli comprises said sine of said measured axial position of said surface rotating around said shaft.

18. The system according to claim 16 further comprising means for measuring the axial position of said surface rotating around said shaft wherein at least one of said stimuli comprises said cosine of said measured axial position of said surface rotating around said shaft.

19. The system according to claim 1 wherein said adaptive controller neural network receives said reference signal and at least one of said stimuli measurements as inputs and produces an adjustment signal.

21. The system according to claim 1 wherein said emulator plant gradient is generated by passing the signals representing the measured noise and vibration back through the emulator network and calculating the gradients at each layer by backpropagating the gradients at each layer of the said emulator back to the input of the neural net controller.

22. The system according to claim 21 wherein said controller neural network is adapted with a gradient descent algorithm based on said plant gradient produced as output by said emulator neural network.

23. The system according to claim 22 wherein said preprocessing filtering module comprises a bandwidth filter.

24. The system according to claim 1 wherein said time-based filter comprises memory delay lines which store the input values for a given time period and permit access to each of the stored values.

25. The system according to claim 24 wherein said state of altering means input signal to said emulator networks comprises said output signal from said controller network for controlling the state of said altering means.

26. The system according to claim 25 wherein said surface is being driven through said fluid.

27. The system according to claim 26 wherein said driven surface is attached to a shaft and driven rotationally through said fluid.

28. The system according to claim 27 further comprising means for measuring the rotational rate at which said shaft is driven wherein said reference signal is comprised of harmonics of the rotational rate at which the surface is driven through the fluid.

29. The system according to claim 28 wherein said harmonics include the harmonic at integer multiples of the rotational rate.

30. The system according to claim 29 wherein said integer multiples include twice the rotational rate.

31. The system according to claim 27 further comprising means for measuring the axial position of said surface rotating around said shaft wherein at least one of said stimuli comprises the axial position of said surface rotating around said shaft.

32. The system according to claim 27 further comprising means for measuring the axial position of said surface rotating around said shaft wherein at least one of said stimuli comprises said sine of said measured axial position of said surface rotating around said shaft.

33. The system according to claim 32 further comprising means for measuring the axial position of said surface rotating around said shaft wherein at least one of said stimuli comprises said cosine of said measured axial position of said surface rotating around said shaft.

34. The system according to claim 32 further comprising means for measuring the forward velocity of a body attached to said surface through said fluid wherein at least one of said stimuli comprises said measured forward velocity.

35. The system according to claim 27 further comprising means for measuring the forward velocity of said surface through said fluid wherein at least one of said stimuli comprises said measured forward velocity.

36. The system according to claim 27 further comprising means for measuring the forward velocity of a body attached to said surface through said fluid wherein at least one of said stimuli comprises said measured forward velocity.

37. The system according to claim 36 further comprising means for measuring the axial position of said surface rotating around said shaft wherein at least one of said stimuli comprises said sine of said measured axial position of said surface rotating around said shaft.

38. The system according to claim 36 further comprising means for measuring the axial position of said surface rotating around said shaft wherein at least one of said stimuli comprises said cosine of said measured axial position of said surface rotating around said shaft.

39. The system according to claim 1 wherein said preprocessing filtering module comprises:

means for digitizing the noise and vibration measurements;

means for extracting targeted signals within a desired frequency bandwidth from the digitized measurements;

means for identifying an envelope of loci of the targeted signal bandwidth and quantifying a single discrete value representing that envelope; and

means for rate-compressing the loci envelope value.

20. The system according to 4 wherein said adaptive controller neural network receives said reference signal and at least one of said stimuli measurements as inputs and produces an adjustment signal.

40. A system for actively controlling the noise and vibration produced by the blades of a rotorcraft as measured by a plurality of noise and vibration sensors mounted on the rotor blades and within the rotorcraft by altering the aerodynamic characteristics of the rotor blades, comprising:

a preprocessing filtering module which isolates the measured noise and vibration desired to be controlled;

a sensor for measuring the rotational rate at which the rotor of the rotorcraft is operating;

means for determining the angular positions of said blades;

a reference signal;

an adaptive feed-forward three-layer controller neural network adapted to receive as input said reference signal and a positional velocity of said rotorcraft and produce an adjustment signal wherein said adjustment signal comprises multiple phase and amplitude shifted harmonics of said reference signal;

means for altering the airflow over said blades;

a feed-forward three-layer emulator neural network trained to model the dynamics of the blade as well as the results and transfer function of changes induced by said altering means wherein said model receives as input the state of said altering means, the positional velocity of said rotorcraft, and the relative positions of said blades, calculates estimated vibration and noise envelope signals, compares the estimated signals with said sensed vibration and noise envelope signals, produces plant gradient signals representing the errors in the emulator model of the plant, and feeds back said plant gradient signals to said controller neural network to allow said controller neural network to adapt itself during runtime using a gradient descent algorithm based on said plant gradient output of said emulator neural network and said sensor preprocessing module; and

a time-based filter which produces time-based representations of the inputs to said emulator neural network and said controller neural network comprising delay lines which store the input values for a given time period and permit access to each of the stored values.

41. The system according to claim 40 wherein said emulator further comprises:

an input vector with dimension J;

a first layer having a number of neurons (Q+1);

an output layer having a number of neurons (L+1) equal to number of sensors for measuring the targeted noise plus the number of sensors for measuring the targeted vibration;

one or more intermediate layers having a number of neurons (M+1);

a set of connections w_j,q⁽¹⁾ between the input vector and the neurons in the first layer, w_q,m⁽²⁾, between the neurons in any successive intermediate layers, and w_m,l⁽³⁾ between the neurons in the last intermediate layer and said output layer;

a modeling section for calculating an estimated result from each of the input, output and intermediate neurons;

a memory section for storing a plurality of prior estimated results from the neurons in said input layer, said output layer and each of said intermediate layers;

a performance section for calculating the gradient of the plant during a backward pass of said measured noise and vibration; and

an update section for update the plant gradients to said controller network.

42. The system according to claim 41 wherein said controller further comprises:

an input vector with dimension J;

a first layer having a number of neurons (K+1);

an output layer having a number of neurons (L+1) equal to number of means for altering the fluid flow;

one or more intermediate layers having a number of neurons (P+1);

a set of connections w_j,k^(1c) between the input vector and the neurons in the input layer and first layer, w_k,p^(2c) between the neurons in any successive intermediate layers, and w_p,l^(3c) between the neurons in the last intermediate layer and said output layer;

a objective function section for calculating an adjustment signal;

a gradient descent performance section for adapting the connections {w_j,k^(1c), w_k,p^(2c),w_p,l^(3c)} according to the plant gradients received from the emulator network; and

an update section for updating the adjustment signal to the altering means and to the input of the emulator network.

43. The system according to claim 42 wherein said blade position is represented by the sine and cosine of the relative angle of said rotor to the body of said rotorcraft.

44. The system according to claim 43 wherein said input vector to said emulator comprises at least one of said blade positions stored in said time-based input filter to said emulator.

45. The system according to claim 44 wherein said input vector to said emulator comprises at least one of said positional velocities stored in said time-based input filter related to said emulator.

46. The system according to claim 45 wherein said input vector to said controller comprises at least one of said positional velocities stored in said time-based input filter related to said controller.

47. The system according to claim 46 wherein said reference signal is comprised of harmonics of said rotor rotational rate.

48. The system according to claim 47 wherein said harmonics include the harmonic at integer multiples of the rotational rate.

49. The system according to claim 40 wherein said input vector to said controller comprises at least one of said positional velocities stored in said time-based input filter to said controller.